Ivan Leo

Leo began in frontend engineering at Credit Suisse in Singapore and learned machine learning through independent study and a paper-reading group. He later worked on Instructor with Jason Liu. He joined Manus on July 27, 2025, building Mail Manus, file suggestions, API integrations and payments support before moving to Google DeepMind.[2][5][1]

Make an evaluation small enough to run often

In Write Stupid Evals, Leo starts with checks a developer can inspect: valid JSON, required fields, retrieval hits and correct tool selection. His text-to-SQL example measures whether the required snippet appears among the retrieved candidates, allowing changes in retrieval quality to be weighed against latency. He then extends single-tool checks to multiple tools and argument validation while retaining the earlier baseline. The aim is a fast local feedback loop that grows alongside the application.[4]

Represent an agent session as an interaction

In his Google DeepMind talk, Leo presents the Interactions API’s server-side state and interaction identifiers, typed multimodal outputs and combined built-in and custom tools. The publisher’s chaptered description also covers Managed Agents: persistent remote sandboxes, environment identifiers, loadable sources and a proxy for injecting credentials. These interfaces organize state and execution around a continuing agent task.[6]

His Manus retrospective treats delivery as work extending beyond a successful deployment. A working email-triggered prototype exposes product questions early; watching colleagues use it reveals confusing interfaces. He describes learning Go while repairing the API’s attachment and connector behavior, and using support-ticket clustering to understand failure modes across languages. These examples connect implementation, documentation, integration testing and user feedback.[5]